AWS makes its EC2 Graviton5 instances available and deploys the formally verified Nitro Isolation Engine

AWS has introduced its M9g and M9gd Amazon EC2 instances, powered by the new Graviton5 processor, alongside the rollout of the Nitro Isolation Engine. This advancement positions the Nitro Hypervisor as the first cloud hypervisor to feature formally verified isolation code between virtual machines, providing a mathematical proof of behavior that surpasses standard testing coverage.

Unveiled at re:Invent 2025, the M9g and M9gd instances are now available to the public. AWS accompanies this launch with two key announcements: one detailing processor performance improvements and the other emphasizing the security assurances afforded by the formal verification process. Over 120,000 customers are already utilizing the Graviton line of processors.

The formal verification of a hypervisor marks a significant milestone in the cloud industry. This method involves mathematically demonstrating that the code functions exactly as specified across all possible system states, rather than just in scenarios covered by traditional testing. Previously limited to critical components, this process now addresses the isolation boundary between different customers’ virtual machines on a shared infrastructure.

Understanding the Scope of the Verification

The Nitro Isolation Engine serves as the sole access point to the memory, processor registers, and input/output devices of virtual machines. Other aspects of the Nitro Hypervisor operate through a restricted interface, with no direct access to customer data. The verification covers four main properties: confidentiality and integrity of virtual machines, the functional correctness of each system call, absence of runtime errors, and memory security. Written in Rust, this trusted computing base relies on approximately 330,000 lines of machine-verified proofs, comparable in scope to the seL4 project that demonstrated the feasibility of verifying realistic operating systems.

However, the scope of the verification must be clearly defined. Currently, the verification covers system calls during the central lifecycle of virtual machines—startup, execution, and shutdown—not the entire hypervisor. Moreover, the proofs depend on explicit assumptions, including the correctness of the Rust compiler and the underlying hardware, as outlined in the accompanying white paper. Thus, the guarantee remains mathematical within a defined perimeter and subject to these assumptions, marking a distinction between verified assurance and universal infallibility.

Shifting Trust Towards Assumptions

For security professionals, this development is significant. While traditional testing samples a subset of scenarios, formal proof encompasses all states within the verified perimeter, enhancing confidence in workload isolation. However, this raises a new question: rather than focusing on whether isolation has been sufficiently tested, attention should shift to whether the underlying assumptions of the proof hold and what aspects remain outside the verified scope.

AWS plans to extend formal verification to other security-impacting components of the Nitro Hypervisor and to make the source code and proofs available for independent review. These initiatives are part of a broader commitment. Launched in 2017, the Nitro System was the first cloud platform without operator access to customer data, with its security assurances validated by an audit from NCC Group.

A Security Benefit That Doesn’t Shift Legal Liabilities

Proven isolation between virtual machines helps mitigate technical risks, specifically the possibility of data leaks between clients on a shared infrastructure. However, it does not alter the legal exposure of processed data. AWS remains a U.S.-based provider, and European data stored on its infrastructure remains subject to extraterritorial regulations, regardless of the mathematical robustness of the hardware isolation. The distinction between technical airtightness and legal permeability is crucial and echoes ongoing discussions around AWS’s offering for European sovereignty.

For organizations subject to sovereignty requirements, the formal verification of the hypervisor represents a noteworthy security argument to be factored into technical evaluations, but it should not be conflated with jurisdictional compliance. The binary nature of sovereignty persists: enhanced technical isolation does not equate to immunity from extraterritorial laws.

Graviton5: A Performance Leap for Workloads

The hardware aspect complements this security framework. The Graviton5 features 192 cores per chip and a level-five cache that is five times larger than its predecessor, with each core offering 2.6 times more cache than Graviton4. AWS claims that DDR5-8800 memory is the fastest available in the cloud, with inter-core latency reduced by up to 33%. The processor boasts up to 25% better performance, with notable improvements of 35% in web applications, 35% in machine learning inference, and 30% in database operations. The M9gd instances also incorporate up to 11.4 TB of local NVMe storage and a 30% increase in input/output operations.

AWS has shared several deployments as examples. Meta has committed to deploying tens of millions of Graviton5 cores for its workloads. Airbnb reports 25% gains on its production search workloads, Atlassian claims 30% better performance and 20% lower latency for Jira versus Graviton4, while SAP HANA Cloud cites enhancements of 35% to 60% on its transactional queries. These figures illustrate the target applications of the M9g and M9gd instances—high-performance computational tasks and scaling workloads.

Scroll to Top